runtime, cmd/internal/obj/arm: improve arm function prologue

When stack growth is not needed, as it usually is not,
execute only a single conditional branch
rather than three conditional instructions.
This adds 4 bytes to every function,
but might speed up execution in the common case.

Sample disassembly for

func f() {
	_ = [128]byte{}
}

Before:

TEXT main.f(SB) x.go
	x.go:3	0x2000	e59a1008	MOVW 0x8(R10), R1
	x.go:3	0x2004	e59fb028	MOVW 0x28(R15), R11
	x.go:3	0x2008	e08d200b	ADD R11, R13, R2
	x.go:3	0x200c	e1520001	CMP R1, R2
	x.go:3	0x2010	91a0300e	MOVW.LS R14, R3
	x.go:3	0x2014	9b0118a9	BL.LS runtime.morestack_noctxt(SB)
	x.go:3	0x2018	9afffff8	B.LS main.f(SB)
	x.go:3	0x201c	e52de084	MOVW.W R14, -0x84(R13)
	x.go:4	0x2020	e28d1004	ADD $4, R13, R1
	x.go:4	0x2024	e3a00000	MOVW $0, R0
	x.go:4	0x2028	eb012255	BL 0x4a984
	x.go:5	0x202c	e49df084	RET #132
	x.go:5	0x2030	eafffffe	B 0x2030
	x.go:5	0x2034	ffffff7c	?

After:

TEXT main.f(SB) x.go
	x.go:3	0x2000	e59a1008	MOVW 0x8(R10), R1
	x.go:3	0x2004	e59fb02c	MOVW 0x2c(R15), R11
	x.go:3	0x2008	e08d200b	ADD R11, R13, R2
	x.go:3	0x200c	e1520001	CMP R1, R2
	x.go:3	0x2010	9a000004	B.LS 0x2028
	x.go:3	0x2014	e52de084	MOVW.W R14, -0x84(R13)
	x.go:4	0x2018	e28d1004	ADD $4, R13, R1
	x.go:4	0x201c	e3a00000	MOVW $0, R0
	x.go:4	0x2020	eb0124dc	BL 0x4b398
	x.go:5	0x2024	e49df084	RET #132
	x.go:5	0x2028	e1a0300e	MOVW R14, R3
	x.go:5	0x202c	eb011b0d	BL runtime.morestack_noctxt(SB)
	x.go:5	0x2030	eafffff2	B main.f(SB)
	x.go:5	0x2034	eafffffe	B 0x2034
	x.go:5	0x2038	ffffff7c	?

Updates #10587.

package sort benchmarks on an iPhone 6:

name            old time/op  new time/op  delta
SortString1K     569µs ± 0%   565µs ± 1%  -0.75%  (p=0.000 n=23+24)
StableString1K   872µs ± 1%   870µs ± 1%  -0.16%  (p=0.009 n=23+24)
SortInt1K        317µs ± 2%   316µs ± 2%    ~     (p=0.410 n=26+26)
StableInt1K      343µs ± 1%   339µs ± 1%  -1.07%  (p=0.000 n=22+23)
SortInt64K      30.0ms ± 1%  30.0ms ± 1%    ~     (p=0.091 n=25+24)
StableInt64K    30.2ms ± 0%  30.0ms ± 0%  -0.69%  (p=0.000 n=22+22)
Sort1e2          147µs ± 1%   146µs ± 0%  -0.48%  (p=0.000 n=25+24)
Stable1e2        290µs ± 1%   286µs ± 1%  -1.30%  (p=0.000 n=23+24)
Sort1e4         29.5ms ± 2%  29.7ms ± 1%  +0.71%  (p=0.000 n=23+23)
Stable1e4       88.7ms ± 4%  88.6ms ± 8%  -0.07%  (p=0.022 n=26+26)
Sort1e6          4.81s ± 7%   4.83s ± 7%    ~     (p=0.192 n=26+26)
Stable1e6        18.3s ± 1%   18.1s ± 1%  -0.76%  (p=0.000 n=25+23)
SearchWrappers   318ns ± 1%   344ns ± 1%  +8.14%  (p=0.000 n=23+26)

package sort benchmarks on a first generation rpi:

name            old time/op  new time/op  delta
SearchWrappers  4.13µs ± 0%  3.95µs ± 0%   -4.42%  (p=0.000 n=15+13)
SortString1K    5.81ms ± 1%  5.82ms ± 2%     ~     (p=0.400 n=14+15)
StableString1K  9.69ms ± 1%  9.73ms ± 0%     ~     (p=0.121 n=15+11)
SortInt1K       3.30ms ± 2%  3.66ms ±19%  +10.82%  (p=0.000 n=15+14)
StableInt1K     5.97ms ±15%  4.17ms ± 8%  -30.05%  (p=0.000 n=15+15)
SortInt64K       319ms ± 1%   295ms ± 1%   -7.65%  (p=0.000 n=15+15)
StableInt64K     343ms ± 0%   332ms ± 0%   -3.26%  (p=0.000 n=12+13)
Sort1e2         3.36ms ± 2%  3.22ms ± 4%   -4.10%  (p=0.000 n=15+15)
Stable1e2       6.74ms ± 1%  6.43ms ± 2%   -4.67%  (p=0.000 n=15+15)
Sort1e4          247ms ± 1%   247ms ± 1%     ~     (p=0.331 n=15+14)
Stable1e4        864ms ± 0%   820ms ± 0%   -5.15%  (p=0.000 n=14+15)
Sort1e6          41.2s ± 0%   41.2s ± 0%   +0.15%  (p=0.000 n=13+14)
Stable1e6         192s ± 0%    182s ± 0%   -5.07%  (p=0.000 n=14+14)

Change-Id: I8a9db77e1d4ea1956575895893bc9d04bd81204b
Reviewed-on: https://go-review.googlesource.com/10497
Reviewed-by: Russ Cox <[email protected]>

Author: josharian

src/cmd/internal/obj/arm/obj5.go

@@ -795,38 +795,45 @@ func stacksplit(ctxt *obj.Link, p *obj.Prog, framesize int32) *obj.Prog {
 		p.Scond = C_SCOND_NE
 	}
 
-	// MOVW.LS	R14, R3
-	p = obj.Appendp(ctxt, p)
-
-	p.As = AMOVW
-	p.Scond = C_SCOND_LS
-	p.From.Type = obj.TYPE_REG
-	p.From.Reg = REGLINK
-	p.To.Type = obj.TYPE_REG
-	p.To.Reg = REG_R3
-
-	// BL.LS		runtime.morestack(SB) // modifies LR, returns with LO still asserted
-	p = obj.Appendp(ctxt, p)
+	// BLS call-to-morestack
+	bls := obj.Appendp(ctxt, p)
+	bls.As = ABLS
+	bls.To.Type = obj.TYPE_BRANCH
 
-	p.As = ABL
-	p.Scond = C_SCOND_LS
-	p.To.Type = obj.TYPE_BRANCH
-	if ctxt.Cursym.Cfunc != 0 {
-		p.To.Sym = obj.Linklookup(ctxt, "runtime.morestackc", 0)
-	} else if ctxt.Cursym.Text.From3.Offset&obj.NEEDCTXT == 0 {
-		p.To.Sym = obj.Linklookup(ctxt, "runtime.morestack_noctxt", 0)
-	} else {
-		p.To.Sym = obj.Linklookup(ctxt, "runtime.morestack", 0)
+	var last *obj.Prog
+	for last = ctxt.Cursym.Text; last.Link != nil; last = last.Link {
 	}
 
-	// BLS	start
-	p = obj.Appendp(ctxt, p)
+	// MOVW	LR, R3
+	movw := obj.Appendp(ctxt, last)
+	movw.As = AMOVW
+	movw.From.Type = obj.TYPE_REG
+	movw.From.Reg = REGLINK
+	movw.To.Type = obj.TYPE_REG
+	movw.To.Reg = REG_R3
+
+	bls.Pcond = movw
+
+	// BL runtime.morestack
+	call := obj.Appendp(ctxt, movw)
+	call.As = obj.ACALL
+	call.To.Type = obj.TYPE_BRANCH
+	morestack := "runtime.morestack"
+	switch {
+	case ctxt.Cursym.Cfunc != 0:
+		morestack = "runtime.morestackc"
+	case ctxt.Cursym.Text.From3.Offset&obj.NEEDCTXT == 0:
+		morestack = "runtime.morestack_noctxt"
+	}
+	call.To.Sym = obj.Linklookup(ctxt, morestack, 0)
 
-	p.As = ABLS
-	p.To.Type = obj.TYPE_BRANCH
-	p.Pcond = ctxt.Cursym.Text.Link
+	// B start
+	b := obj.Appendp(ctxt, call)
+	b.As = obj.AJMP
+	b.To.Type = obj.TYPE_BRANCH
+	b.Pcond = ctxt.Cursym.Text.Link
 
-	return p
+	return bls
 }
 
 func initdiv(ctxt *obj.Link) {

src/runtime/asm_arm.s

@@ -259,7 +259,6 @@ noswitch:
 
 // Called during function prolog when more stack is needed.
 // R1 frame size
-// R2 arg size
 // R3 prolog's LR
 // NB. we do not save R0 because we've forced 5c to pass all arguments
 // on the stack.

src/runtime/sys_arm.go

@@ -24,7 +24,7 @@ func rewindmorestack(buf *gobuf) {
 	var inst uint32
 	if buf.pc&3 == 0 && buf.pc != 0 {
 		inst = *(*uint32)(unsafe.Pointer(buf.pc))
-		if inst>>24 == 0x9a {
+		if inst>>24 == 0x9a || inst>>24 == 0xea {
 			buf.pc += uintptr(int32(inst<<8)>>6) + 8
 			return
 		}